EDI & AI policies

Science is a collective effort. It works best when people with different backgrounds, experiences, perspectives, and ways of thinking can contribute fully.

In our lab, equality, diversity, and inclusion are not distinct from our research values. Much of our work studies how bias can emerge, transfer, and amplify in learning systems, including through choices that may look neutral at first. We take the same lesson seriously in how we work together: exclusion, unfairness, and unequal opportunity often arise through systems, habits, and assumptions unless we actively question them.

We want the lab to be a place where people can do careful, creative science while being treated with respect. We commit to:

• Welcoming and supporting people across backgrounds, identities, disciplines, nationalities, career stages, personal circumstances, and ways of thinking.
• Treating lab members, collaborators, students, and visitors with respect, and taking bias, discrimination, harassment, and exclusion seriously.
• Making expectations, opportunities, authorship, and credit as transparent as possible.
• Creating space for people to raise concerns, ask questions, suggest improvements, and contribute to the lab culture.
• Remembering that fairness is not only something we study in models, but something we practice in our community.

Our lab is not only its scientific output. The quality and impact of our research depend on the people doing it, the environment we create together, and our willingness to keep improving.

This statement is a living document. We will revisit it as the group grows and as we learn better ways to support one another.

AI tools, including large language models and coding agents, are useful research tools. Used well, they can speed up coding, writing, analysis, and project organization. Used carelessly, they can introduce plausible-looking mistakes, hide weak reasoning, or make our work less transparent.

The main principle is simple: you are responsible for everything you produce, regardless of the tools used (our policy is inspired by the AIME Agents/LLM Use Policy).

Principles

1. Use AI as a tool

AI tools can help with brainstorming, drafting, editing, coding, debugging, explaining concepts, summarising notes, and making routine work faster. They are not a substitute for scientific judgment, domain knowledge, careful reading, or verification.

2. Be transparent

Be open with the group about where and how you used AI tools when the use is relevant to the work.

Examples of relevant use include:

• AI materially helped write or restructure text that will be shared outside the group.
• AI generated, rewrote, or debugged code used in a project.
• AI helped with analysis, coding methods, generating data, or interpreting results.
• An agent took actions on your behalf, such as editing files, running commands, or preparing a larger change.

3. Be accountable

Whatever you produce is your responsibility. This applies to text, code, figures, analyses, citations, and conclusions. If an AI-assisted output contains an error, the error is still yours to catch and correct.

Required Practice

References

Do not cite references that you have not read, and do not trust AI-generated references. AI can help identify relevant references that we were not aware of, but it is our responsibility to read them before citing them.

Verify the output

Before sharing or relying on AI-assisted work:

• Read the output carefully.
• Check factual claims against reliable sources.
• Run and test code yourself.
• Inspect code changes, especially if an agent edited multiple files.
• Confirm that the result answers the actual question or task.

Follow disclosure rules

If a venue, journal, or collaborator requires AI-use disclosure, follow that rule.

Protect confidential material

Do not paste confidential, private, or sensitive material into hosted AI tools unless you know that the tool and the data-sharing agreement allow it.

This includes:

• Unpublished work from collaborators.
• Peer-review material.
• Personal data.
• Credentials, tokens, SSH keys, or private URLs.

Keep work reproducible

If AI use materially affects a result, record enough information for future you and collaborators to understand what happened. Depending on the case, this may include the tool, model, date, prompt, generated output, edited code, command log, or a short note in a methods file or lab notebook.

What to Do When Unsure?

Ask the group, your supervisor, or the relevant collaborator before using AI tools.